3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'slab_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/proc_fs.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/uaccess.h>
108 #include <linux/nodemask.h>
109 #include <linux/kmemleak.h>
110 #include <linux/mempolicy.h>
111 #include <linux/mutex.h>
112 #include <linux/fault-inject.h>
113 #include <linux/rtmutex.h>
114 #include <linux/reciprocal_div.h>
115 #include <linux/debugobjects.h>
116 #include <linux/kmemcheck.h>
117 #include <linux/memory.h>
118 #include <linux/prefetch.h>
120 #include <net/sock.h>
122 #include <asm/cacheflush.h>
123 #include <asm/tlbflush.h>
124 #include <asm/page.h>
126 #include <trace/events/kmem.h>
128 #include "internal.h"
133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
134 * 0 for faster, smaller code (especially in the critical paths).
136 * STATS - 1 to collect stats for /proc/slabinfo.
137 * 0 for faster, smaller code (especially in the critical paths).
139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
142 #ifdef CONFIG_DEBUG_SLAB
145 #define FORCED_DEBUG 1
149 #define FORCED_DEBUG 0
152 /* Shouldn't this be in a header file somewhere? */
153 #define BYTES_PER_WORD sizeof(void *)
154 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
156 #ifndef ARCH_KMALLOC_FLAGS
157 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
160 #define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
161 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
163 #if FREELIST_BYTE_INDEX
164 typedef unsigned char freelist_idx_t
;
166 typedef unsigned short freelist_idx_t
;
169 #define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
175 * - LIFO ordering, to hand out cache-warm objects from _alloc
176 * - reduce the number of linked list operations
177 * - reduce spinlock operations
179 * The limit is stored in the per-cpu structure to reduce the data cache
186 unsigned int batchcount
;
187 unsigned int touched
;
189 * Must have this definition in here for the proper
190 * alignment of array_cache. Also simplifies accessing
197 struct array_cache ac
;
201 * Need this for bootstrapping a per node allocator.
203 #define NUM_INIT_LISTS (2 * MAX_NUMNODES)
204 static struct kmem_cache_node __initdata init_kmem_cache_node
[NUM_INIT_LISTS
];
205 #define CACHE_CACHE 0
206 #define SIZE_NODE (MAX_NUMNODES)
208 static int drain_freelist(struct kmem_cache
*cache
,
209 struct kmem_cache_node
*n
, int tofree
);
210 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
211 int node
, struct list_head
*list
);
212 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
);
213 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
214 static void cache_reap(struct work_struct
*unused
);
216 static int slab_early_init
= 1;
218 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
220 static void kmem_cache_node_init(struct kmem_cache_node
*parent
)
222 INIT_LIST_HEAD(&parent
->slabs_full
);
223 INIT_LIST_HEAD(&parent
->slabs_partial
);
224 INIT_LIST_HEAD(&parent
->slabs_free
);
225 parent
->shared
= NULL
;
226 parent
->alien
= NULL
;
227 parent
->colour_next
= 0;
228 spin_lock_init(&parent
->list_lock
);
229 parent
->free_objects
= 0;
230 parent
->free_touched
= 0;
233 #define MAKE_LIST(cachep, listp, slab, nodeid) \
235 INIT_LIST_HEAD(listp); \
236 list_splice(&get_node(cachep, nodeid)->slab, listp); \
239 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
241 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
242 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
243 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
246 #define CFLGS_OBJFREELIST_SLAB (0x40000000UL)
247 #define CFLGS_OFF_SLAB (0x80000000UL)
248 #define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB)
249 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
251 #define BATCHREFILL_LIMIT 16
253 * Optimization question: fewer reaps means less probability for unnessary
254 * cpucache drain/refill cycles.
256 * OTOH the cpuarrays can contain lots of objects,
257 * which could lock up otherwise freeable slabs.
259 #define REAPTIMEOUT_AC (2*HZ)
260 #define REAPTIMEOUT_NODE (4*HZ)
263 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
264 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
265 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
266 #define STATS_INC_GROWN(x) ((x)->grown++)
267 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
268 #define STATS_SET_HIGH(x) \
270 if ((x)->num_active > (x)->high_mark) \
271 (x)->high_mark = (x)->num_active; \
273 #define STATS_INC_ERR(x) ((x)->errors++)
274 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
275 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
276 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
277 #define STATS_SET_FREEABLE(x, i) \
279 if ((x)->max_freeable < i) \
280 (x)->max_freeable = i; \
282 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
283 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
284 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
285 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
287 #define STATS_INC_ACTIVE(x) do { } while (0)
288 #define STATS_DEC_ACTIVE(x) do { } while (0)
289 #define STATS_INC_ALLOCED(x) do { } while (0)
290 #define STATS_INC_GROWN(x) do { } while (0)
291 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
292 #define STATS_SET_HIGH(x) do { } while (0)
293 #define STATS_INC_ERR(x) do { } while (0)
294 #define STATS_INC_NODEALLOCS(x) do { } while (0)
295 #define STATS_INC_NODEFREES(x) do { } while (0)
296 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
297 #define STATS_SET_FREEABLE(x, i) do { } while (0)
298 #define STATS_INC_ALLOCHIT(x) do { } while (0)
299 #define STATS_INC_ALLOCMISS(x) do { } while (0)
300 #define STATS_INC_FREEHIT(x) do { } while (0)
301 #define STATS_INC_FREEMISS(x) do { } while (0)
307 * memory layout of objects:
309 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
310 * the end of an object is aligned with the end of the real
311 * allocation. Catches writes behind the end of the allocation.
312 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
314 * cachep->obj_offset: The real object.
315 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
316 * cachep->size - 1* BYTES_PER_WORD: last caller address
317 * [BYTES_PER_WORD long]
319 static int obj_offset(struct kmem_cache
*cachep
)
321 return cachep
->obj_offset
;
324 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
326 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
327 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
328 sizeof(unsigned long long));
331 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
333 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
334 if (cachep
->flags
& SLAB_STORE_USER
)
335 return (unsigned long long *)(objp
+ cachep
->size
-
336 sizeof(unsigned long long) -
338 return (unsigned long long *) (objp
+ cachep
->size
-
339 sizeof(unsigned long long));
342 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
344 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
345 return (void **)(objp
+ cachep
->size
- BYTES_PER_WORD
);
350 #define obj_offset(x) 0
351 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
352 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
353 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
357 #ifdef CONFIG_DEBUG_SLAB_LEAK
359 static inline bool is_store_user_clean(struct kmem_cache
*cachep
)
361 return atomic_read(&cachep
->store_user_clean
) == 1;
364 static inline void set_store_user_clean(struct kmem_cache
*cachep
)
366 atomic_set(&cachep
->store_user_clean
, 1);
369 static inline void set_store_user_dirty(struct kmem_cache
*cachep
)
371 if (is_store_user_clean(cachep
))
372 atomic_set(&cachep
->store_user_clean
, 0);
376 static inline void set_store_user_dirty(struct kmem_cache
*cachep
) {}
381 * Do not go above this order unless 0 objects fit into the slab or
382 * overridden on the command line.
384 #define SLAB_MAX_ORDER_HI 1
385 #define SLAB_MAX_ORDER_LO 0
386 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
387 static bool slab_max_order_set __initdata
;
389 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
391 struct page
*page
= virt_to_head_page(obj
);
392 return page
->slab_cache
;
395 static inline void *index_to_obj(struct kmem_cache
*cache
, struct page
*page
,
398 return page
->s_mem
+ cache
->size
* idx
;
402 * We want to avoid an expensive divide : (offset / cache->size)
403 * Using the fact that size is a constant for a particular cache,
404 * we can replace (offset / cache->size) by
405 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
407 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
408 const struct page
*page
, void *obj
)
410 u32 offset
= (obj
- page
->s_mem
);
411 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
414 #define BOOT_CPUCACHE_ENTRIES 1
415 /* internal cache of cache description objs */
416 static struct kmem_cache kmem_cache_boot
= {
418 .limit
= BOOT_CPUCACHE_ENTRIES
,
420 .size
= sizeof(struct kmem_cache
),
421 .name
= "kmem_cache",
424 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
426 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
428 return this_cpu_ptr(cachep
->cpu_cache
);
432 * Calculate the number of objects and left-over bytes for a given buffer size.
434 static unsigned int cache_estimate(unsigned long gfporder
, size_t buffer_size
,
435 unsigned long flags
, size_t *left_over
)
438 size_t slab_size
= PAGE_SIZE
<< gfporder
;
441 * The slab management structure can be either off the slab or
442 * on it. For the latter case, the memory allocated for a
445 * - @buffer_size bytes for each object
446 * - One freelist_idx_t for each object
448 * We don't need to consider alignment of freelist because
449 * freelist will be at the end of slab page. The objects will be
450 * at the correct alignment.
452 * If the slab management structure is off the slab, then the
453 * alignment will already be calculated into the size. Because
454 * the slabs are all pages aligned, the objects will be at the
455 * correct alignment when allocated.
457 if (flags
& (CFLGS_OBJFREELIST_SLAB
| CFLGS_OFF_SLAB
)) {
458 num
= slab_size
/ buffer_size
;
459 *left_over
= slab_size
% buffer_size
;
461 num
= slab_size
/ (buffer_size
+ sizeof(freelist_idx_t
));
462 *left_over
= slab_size
%
463 (buffer_size
+ sizeof(freelist_idx_t
));
470 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
472 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
475 pr_err("slab error in %s(): cache `%s': %s\n",
476 function
, cachep
->name
, msg
);
478 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
483 * By default on NUMA we use alien caches to stage the freeing of
484 * objects allocated from other nodes. This causes massive memory
485 * inefficiencies when using fake NUMA setup to split memory into a
486 * large number of small nodes, so it can be disabled on the command
490 static int use_alien_caches __read_mostly
= 1;
491 static int __init
noaliencache_setup(char *s
)
493 use_alien_caches
= 0;
496 __setup("noaliencache", noaliencache_setup
);
498 static int __init
slab_max_order_setup(char *str
)
500 get_option(&str
, &slab_max_order
);
501 slab_max_order
= slab_max_order
< 0 ? 0 :
502 min(slab_max_order
, MAX_ORDER
- 1);
503 slab_max_order_set
= true;
507 __setup("slab_max_order=", slab_max_order_setup
);
511 * Special reaping functions for NUMA systems called from cache_reap().
512 * These take care of doing round robin flushing of alien caches (containing
513 * objects freed on different nodes from which they were allocated) and the
514 * flushing of remote pcps by calling drain_node_pages.
516 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
518 static void init_reap_node(int cpu
)
522 node
= next_node(cpu_to_mem(cpu
), node_online_map
);
523 if (node
== MAX_NUMNODES
)
524 node
= first_node(node_online_map
);
526 per_cpu(slab_reap_node
, cpu
) = node
;
529 static void next_reap_node(void)
531 int node
= __this_cpu_read(slab_reap_node
);
533 node
= next_node(node
, node_online_map
);
534 if (unlikely(node
>= MAX_NUMNODES
))
535 node
= first_node(node_online_map
);
536 __this_cpu_write(slab_reap_node
, node
);
540 #define init_reap_node(cpu) do { } while (0)
541 #define next_reap_node(void) do { } while (0)
545 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
546 * via the workqueue/eventd.
547 * Add the CPU number into the expiration time to minimize the possibility of
548 * the CPUs getting into lockstep and contending for the global cache chain
551 static void start_cpu_timer(int cpu
)
553 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
556 * When this gets called from do_initcalls via cpucache_init(),
557 * init_workqueues() has already run, so keventd will be setup
560 if (keventd_up() && reap_work
->work
.func
== NULL
) {
562 INIT_DEFERRABLE_WORK(reap_work
, cache_reap
);
563 schedule_delayed_work_on(cpu
, reap_work
,
564 __round_jiffies_relative(HZ
, cpu
));
568 static void init_arraycache(struct array_cache
*ac
, int limit
, int batch
)
571 * The array_cache structures contain pointers to free object.
572 * However, when such objects are allocated or transferred to another
573 * cache the pointers are not cleared and they could be counted as
574 * valid references during a kmemleak scan. Therefore, kmemleak must
575 * not scan such objects.
577 kmemleak_no_scan(ac
);
581 ac
->batchcount
= batch
;
586 static struct array_cache
*alloc_arraycache(int node
, int entries
,
587 int batchcount
, gfp_t gfp
)
589 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
590 struct array_cache
*ac
= NULL
;
592 ac
= kmalloc_node(memsize
, gfp
, node
);
593 init_arraycache(ac
, entries
, batchcount
);
597 static noinline
void cache_free_pfmemalloc(struct kmem_cache
*cachep
,
598 struct page
*page
, void *objp
)
600 struct kmem_cache_node
*n
;
604 page_node
= page_to_nid(page
);
605 n
= get_node(cachep
, page_node
);
607 spin_lock(&n
->list_lock
);
608 free_block(cachep
, &objp
, 1, page_node
, &list
);
609 spin_unlock(&n
->list_lock
);
611 slabs_destroy(cachep
, &list
);
615 * Transfer objects in one arraycache to another.
616 * Locking must be handled by the caller.
618 * Return the number of entries transferred.
620 static int transfer_objects(struct array_cache
*to
,
621 struct array_cache
*from
, unsigned int max
)
623 /* Figure out how many entries to transfer */
624 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
629 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
639 #define drain_alien_cache(cachep, alien) do { } while (0)
640 #define reap_alien(cachep, n) do { } while (0)
642 static inline struct alien_cache
**alloc_alien_cache(int node
,
643 int limit
, gfp_t gfp
)
648 static inline void free_alien_cache(struct alien_cache
**ac_ptr
)
652 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
657 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
663 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
664 gfp_t flags
, int nodeid
)
669 static inline gfp_t
gfp_exact_node(gfp_t flags
)
671 return flags
& ~__GFP_NOFAIL
;
674 #else /* CONFIG_NUMA */
676 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
677 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
679 static struct alien_cache
*__alloc_alien_cache(int node
, int entries
,
680 int batch
, gfp_t gfp
)
682 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct alien_cache
);
683 struct alien_cache
*alc
= NULL
;
685 alc
= kmalloc_node(memsize
, gfp
, node
);
686 init_arraycache(&alc
->ac
, entries
, batch
);
687 spin_lock_init(&alc
->lock
);
691 static struct alien_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
693 struct alien_cache
**alc_ptr
;
694 size_t memsize
= sizeof(void *) * nr_node_ids
;
699 alc_ptr
= kzalloc_node(memsize
, gfp
, node
);
704 if (i
== node
|| !node_online(i
))
706 alc_ptr
[i
] = __alloc_alien_cache(node
, limit
, 0xbaadf00d, gfp
);
708 for (i
--; i
>= 0; i
--)
717 static void free_alien_cache(struct alien_cache
**alc_ptr
)
728 static void __drain_alien_cache(struct kmem_cache
*cachep
,
729 struct array_cache
*ac
, int node
,
730 struct list_head
*list
)
732 struct kmem_cache_node
*n
= get_node(cachep
, node
);
735 spin_lock(&n
->list_lock
);
737 * Stuff objects into the remote nodes shared array first.
738 * That way we could avoid the overhead of putting the objects
739 * into the free lists and getting them back later.
742 transfer_objects(n
->shared
, ac
, ac
->limit
);
744 free_block(cachep
, ac
->entry
, ac
->avail
, node
, list
);
746 spin_unlock(&n
->list_lock
);
751 * Called from cache_reap() to regularly drain alien caches round robin.
753 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
)
755 int node
= __this_cpu_read(slab_reap_node
);
758 struct alien_cache
*alc
= n
->alien
[node
];
759 struct array_cache
*ac
;
763 if (ac
->avail
&& spin_trylock_irq(&alc
->lock
)) {
766 __drain_alien_cache(cachep
, ac
, node
, &list
);
767 spin_unlock_irq(&alc
->lock
);
768 slabs_destroy(cachep
, &list
);
774 static void drain_alien_cache(struct kmem_cache
*cachep
,
775 struct alien_cache
**alien
)
778 struct alien_cache
*alc
;
779 struct array_cache
*ac
;
782 for_each_online_node(i
) {
788 spin_lock_irqsave(&alc
->lock
, flags
);
789 __drain_alien_cache(cachep
, ac
, i
, &list
);
790 spin_unlock_irqrestore(&alc
->lock
, flags
);
791 slabs_destroy(cachep
, &list
);
796 static int __cache_free_alien(struct kmem_cache
*cachep
, void *objp
,
797 int node
, int page_node
)
799 struct kmem_cache_node
*n
;
800 struct alien_cache
*alien
= NULL
;
801 struct array_cache
*ac
;
804 n
= get_node(cachep
, node
);
805 STATS_INC_NODEFREES(cachep
);
806 if (n
->alien
&& n
->alien
[page_node
]) {
807 alien
= n
->alien
[page_node
];
809 spin_lock(&alien
->lock
);
810 if (unlikely(ac
->avail
== ac
->limit
)) {
811 STATS_INC_ACOVERFLOW(cachep
);
812 __drain_alien_cache(cachep
, ac
, page_node
, &list
);
814 ac
->entry
[ac
->avail
++] = objp
;
815 spin_unlock(&alien
->lock
);
816 slabs_destroy(cachep
, &list
);
818 n
= get_node(cachep
, page_node
);
819 spin_lock(&n
->list_lock
);
820 free_block(cachep
, &objp
, 1, page_node
, &list
);
821 spin_unlock(&n
->list_lock
);
822 slabs_destroy(cachep
, &list
);
827 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
829 int page_node
= page_to_nid(virt_to_page(objp
));
830 int node
= numa_mem_id();
832 * Make sure we are not freeing a object from another node to the array
835 if (likely(node
== page_node
))
838 return __cache_free_alien(cachep
, objp
, node
, page_node
);
842 * Construct gfp mask to allocate from a specific node but do not reclaim or
843 * warn about failures.
845 static inline gfp_t
gfp_exact_node(gfp_t flags
)
847 return (flags
| __GFP_THISNODE
| __GFP_NOWARN
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
852 * Allocates and initializes node for a node on each slab cache, used for
853 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
854 * will be allocated off-node since memory is not yet online for the new node.
855 * When hotplugging memory or a cpu, existing node are not replaced if
858 * Must hold slab_mutex.
860 static int init_cache_node_node(int node
)
862 struct kmem_cache
*cachep
;
863 struct kmem_cache_node
*n
;
864 const size_t memsize
= sizeof(struct kmem_cache_node
);
866 list_for_each_entry(cachep
, &slab_caches
, list
) {
868 * Set up the kmem_cache_node for cpu before we can
869 * begin anything. Make sure some other cpu on this
870 * node has not already allocated this
872 n
= get_node(cachep
, node
);
874 n
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
877 kmem_cache_node_init(n
);
878 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
879 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
882 * The kmem_cache_nodes don't come and go as CPUs
883 * come and go. slab_mutex is sufficient
886 cachep
->node
[node
] = n
;
889 spin_lock_irq(&n
->list_lock
);
891 (1 + nr_cpus_node(node
)) *
892 cachep
->batchcount
+ cachep
->num
;
893 spin_unlock_irq(&n
->list_lock
);
898 static void cpuup_canceled(long cpu
)
900 struct kmem_cache
*cachep
;
901 struct kmem_cache_node
*n
= NULL
;
902 int node
= cpu_to_mem(cpu
);
903 const struct cpumask
*mask
= cpumask_of_node(node
);
905 list_for_each_entry(cachep
, &slab_caches
, list
) {
906 struct array_cache
*nc
;
907 struct array_cache
*shared
;
908 struct alien_cache
**alien
;
911 n
= get_node(cachep
, node
);
915 spin_lock_irq(&n
->list_lock
);
917 /* Free limit for this kmem_cache_node */
918 n
->free_limit
-= cachep
->batchcount
;
920 /* cpu is dead; no one can alloc from it. */
921 nc
= per_cpu_ptr(cachep
->cpu_cache
, cpu
);
923 free_block(cachep
, nc
->entry
, nc
->avail
, node
, &list
);
927 if (!cpumask_empty(mask
)) {
928 spin_unlock_irq(&n
->list_lock
);
934 free_block(cachep
, shared
->entry
,
935 shared
->avail
, node
, &list
);
942 spin_unlock_irq(&n
->list_lock
);
946 drain_alien_cache(cachep
, alien
);
947 free_alien_cache(alien
);
951 slabs_destroy(cachep
, &list
);
954 * In the previous loop, all the objects were freed to
955 * the respective cache's slabs, now we can go ahead and
956 * shrink each nodelist to its limit.
958 list_for_each_entry(cachep
, &slab_caches
, list
) {
959 n
= get_node(cachep
, node
);
962 drain_freelist(cachep
, n
, INT_MAX
);
966 static int cpuup_prepare(long cpu
)
968 struct kmem_cache
*cachep
;
969 struct kmem_cache_node
*n
= NULL
;
970 int node
= cpu_to_mem(cpu
);
974 * We need to do this right in the beginning since
975 * alloc_arraycache's are going to use this list.
976 * kmalloc_node allows us to add the slab to the right
977 * kmem_cache_node and not this cpu's kmem_cache_node
979 err
= init_cache_node_node(node
);
984 * Now we can go ahead with allocating the shared arrays and
987 list_for_each_entry(cachep
, &slab_caches
, list
) {
988 struct array_cache
*shared
= NULL
;
989 struct alien_cache
**alien
= NULL
;
991 if (cachep
->shared
) {
992 shared
= alloc_arraycache(node
,
993 cachep
->shared
* cachep
->batchcount
,
994 0xbaadf00d, GFP_KERNEL
);
998 if (use_alien_caches
) {
999 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1005 n
= get_node(cachep
, node
);
1008 spin_lock_irq(&n
->list_lock
);
1011 * We are serialised from CPU_DEAD or
1012 * CPU_UP_CANCELLED by the cpucontrol lock
1023 spin_unlock_irq(&n
->list_lock
);
1025 free_alien_cache(alien
);
1030 cpuup_canceled(cpu
);
1034 static int cpuup_callback(struct notifier_block
*nfb
,
1035 unsigned long action
, void *hcpu
)
1037 long cpu
= (long)hcpu
;
1041 case CPU_UP_PREPARE
:
1042 case CPU_UP_PREPARE_FROZEN
:
1043 mutex_lock(&slab_mutex
);
1044 err
= cpuup_prepare(cpu
);
1045 mutex_unlock(&slab_mutex
);
1048 case CPU_ONLINE_FROZEN
:
1049 start_cpu_timer(cpu
);
1051 #ifdef CONFIG_HOTPLUG_CPU
1052 case CPU_DOWN_PREPARE
:
1053 case CPU_DOWN_PREPARE_FROZEN
:
1055 * Shutdown cache reaper. Note that the slab_mutex is
1056 * held so that if cache_reap() is invoked it cannot do
1057 * anything expensive but will only modify reap_work
1058 * and reschedule the timer.
1060 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1061 /* Now the cache_reaper is guaranteed to be not running. */
1062 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1064 case CPU_DOWN_FAILED
:
1065 case CPU_DOWN_FAILED_FROZEN
:
1066 start_cpu_timer(cpu
);
1069 case CPU_DEAD_FROZEN
:
1071 * Even if all the cpus of a node are down, we don't free the
1072 * kmem_cache_node of any cache. This to avoid a race between
1073 * cpu_down, and a kmalloc allocation from another cpu for
1074 * memory from the node of the cpu going down. The node
1075 * structure is usually allocated from kmem_cache_create() and
1076 * gets destroyed at kmem_cache_destroy().
1080 case CPU_UP_CANCELED
:
1081 case CPU_UP_CANCELED_FROZEN
:
1082 mutex_lock(&slab_mutex
);
1083 cpuup_canceled(cpu
);
1084 mutex_unlock(&slab_mutex
);
1087 return notifier_from_errno(err
);
1090 static struct notifier_block cpucache_notifier
= {
1091 &cpuup_callback
, NULL
, 0
1094 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1096 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1097 * Returns -EBUSY if all objects cannot be drained so that the node is not
1100 * Must hold slab_mutex.
1102 static int __meminit
drain_cache_node_node(int node
)
1104 struct kmem_cache
*cachep
;
1107 list_for_each_entry(cachep
, &slab_caches
, list
) {
1108 struct kmem_cache_node
*n
;
1110 n
= get_node(cachep
, node
);
1114 drain_freelist(cachep
, n
, INT_MAX
);
1116 if (!list_empty(&n
->slabs_full
) ||
1117 !list_empty(&n
->slabs_partial
)) {
1125 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1126 unsigned long action
, void *arg
)
1128 struct memory_notify
*mnb
= arg
;
1132 nid
= mnb
->status_change_nid
;
1137 case MEM_GOING_ONLINE
:
1138 mutex_lock(&slab_mutex
);
1139 ret
= init_cache_node_node(nid
);
1140 mutex_unlock(&slab_mutex
);
1142 case MEM_GOING_OFFLINE
:
1143 mutex_lock(&slab_mutex
);
1144 ret
= drain_cache_node_node(nid
);
1145 mutex_unlock(&slab_mutex
);
1149 case MEM_CANCEL_ONLINE
:
1150 case MEM_CANCEL_OFFLINE
:
1154 return notifier_from_errno(ret
);
1156 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1159 * swap the static kmem_cache_node with kmalloced memory
1161 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_cache_node
*list
,
1164 struct kmem_cache_node
*ptr
;
1166 ptr
= kmalloc_node(sizeof(struct kmem_cache_node
), GFP_NOWAIT
, nodeid
);
1169 memcpy(ptr
, list
, sizeof(struct kmem_cache_node
));
1171 * Do not assume that spinlocks can be initialized via memcpy:
1173 spin_lock_init(&ptr
->list_lock
);
1175 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1176 cachep
->node
[nodeid
] = ptr
;
1180 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1181 * size of kmem_cache_node.
1183 static void __init
set_up_node(struct kmem_cache
*cachep
, int index
)
1187 for_each_online_node(node
) {
1188 cachep
->node
[node
] = &init_kmem_cache_node
[index
+ node
];
1189 cachep
->node
[node
]->next_reap
= jiffies
+
1191 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1196 * Initialisation. Called after the page allocator have been initialised and
1197 * before smp_init().
1199 void __init
kmem_cache_init(void)
1203 BUILD_BUG_ON(sizeof(((struct page
*)NULL
)->lru
) <
1204 sizeof(struct rcu_head
));
1205 kmem_cache
= &kmem_cache_boot
;
1207 if (!IS_ENABLED(CONFIG_NUMA
) || num_possible_nodes() == 1)
1208 use_alien_caches
= 0;
1210 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1211 kmem_cache_node_init(&init_kmem_cache_node
[i
]);
1214 * Fragmentation resistance on low memory - only use bigger
1215 * page orders on machines with more than 32MB of memory if
1216 * not overridden on the command line.
1218 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1219 slab_max_order
= SLAB_MAX_ORDER_HI
;
1221 /* Bootstrap is tricky, because several objects are allocated
1222 * from caches that do not exist yet:
1223 * 1) initialize the kmem_cache cache: it contains the struct
1224 * kmem_cache structures of all caches, except kmem_cache itself:
1225 * kmem_cache is statically allocated.
1226 * Initially an __init data area is used for the head array and the
1227 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1228 * array at the end of the bootstrap.
1229 * 2) Create the first kmalloc cache.
1230 * The struct kmem_cache for the new cache is allocated normally.
1231 * An __init data area is used for the head array.
1232 * 3) Create the remaining kmalloc caches, with minimally sized
1234 * 4) Replace the __init data head arrays for kmem_cache and the first
1235 * kmalloc cache with kmalloc allocated arrays.
1236 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1237 * the other cache's with kmalloc allocated memory.
1238 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1241 /* 1) create the kmem_cache */
1244 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1246 create_boot_cache(kmem_cache
, "kmem_cache",
1247 offsetof(struct kmem_cache
, node
) +
1248 nr_node_ids
* sizeof(struct kmem_cache_node
*),
1249 SLAB_HWCACHE_ALIGN
);
1250 list_add(&kmem_cache
->list
, &slab_caches
);
1251 slab_state
= PARTIAL
;
1254 * Initialize the caches that provide memory for the kmem_cache_node
1255 * structures first. Without this, further allocations will bug.
1257 kmalloc_caches
[INDEX_NODE
] = create_kmalloc_cache("kmalloc-node",
1258 kmalloc_size(INDEX_NODE
), ARCH_KMALLOC_FLAGS
);
1259 slab_state
= PARTIAL_NODE
;
1260 setup_kmalloc_cache_index_table();
1262 slab_early_init
= 0;
1264 /* 5) Replace the bootstrap kmem_cache_node */
1268 for_each_online_node(nid
) {
1269 init_list(kmem_cache
, &init_kmem_cache_node
[CACHE_CACHE
+ nid
], nid
);
1271 init_list(kmalloc_caches
[INDEX_NODE
],
1272 &init_kmem_cache_node
[SIZE_NODE
+ nid
], nid
);
1276 create_kmalloc_caches(ARCH_KMALLOC_FLAGS
);
1279 void __init
kmem_cache_init_late(void)
1281 struct kmem_cache
*cachep
;
1285 /* 6) resize the head arrays to their final sizes */
1286 mutex_lock(&slab_mutex
);
1287 list_for_each_entry(cachep
, &slab_caches
, list
)
1288 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1290 mutex_unlock(&slab_mutex
);
1296 * Register a cpu startup notifier callback that initializes
1297 * cpu_cache_get for all new cpus
1299 register_cpu_notifier(&cpucache_notifier
);
1303 * Register a memory hotplug callback that initializes and frees
1306 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1310 * The reap timers are started later, with a module init call: That part
1311 * of the kernel is not yet operational.
1315 static int __init
cpucache_init(void)
1320 * Register the timers that return unneeded pages to the page allocator
1322 for_each_online_cpu(cpu
)
1323 start_cpu_timer(cpu
);
1329 __initcall(cpucache_init
);
1331 static noinline
void
1332 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1335 struct kmem_cache_node
*n
;
1337 unsigned long flags
;
1339 static DEFINE_RATELIMIT_STATE(slab_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
1340 DEFAULT_RATELIMIT_BURST
);
1342 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slab_oom_rs
))
1345 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1346 nodeid
, gfpflags
, &gfpflags
);
1347 pr_warn(" cache: %s, object size: %d, order: %d\n",
1348 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1350 for_each_kmem_cache_node(cachep
, node
, n
) {
1351 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1352 unsigned long active_slabs
= 0, num_slabs
= 0;
1354 spin_lock_irqsave(&n
->list_lock
, flags
);
1355 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
1356 active_objs
+= cachep
->num
;
1359 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
1360 active_objs
+= page
->active
;
1363 list_for_each_entry(page
, &n
->slabs_free
, lru
)
1366 free_objects
+= n
->free_objects
;
1367 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1369 num_slabs
+= active_slabs
;
1370 num_objs
= num_slabs
* cachep
->num
;
1371 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1372 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1379 * Interface to system's page allocator. No need to hold the
1380 * kmem_cache_node ->list_lock.
1382 * If we requested dmaable memory, we will get it. Even if we
1383 * did not request dmaable memory, we might get it, but that
1384 * would be relatively rare and ignorable.
1386 static struct page
*kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
,
1392 flags
|= cachep
->allocflags
;
1393 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1394 flags
|= __GFP_RECLAIMABLE
;
1396 page
= __alloc_pages_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1398 slab_out_of_memory(cachep
, flags
, nodeid
);
1402 if (memcg_charge_slab(page
, flags
, cachep
->gfporder
, cachep
)) {
1403 __free_pages(page
, cachep
->gfporder
);
1407 nr_pages
= (1 << cachep
->gfporder
);
1408 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1409 add_zone_page_state(page_zone(page
),
1410 NR_SLAB_RECLAIMABLE
, nr_pages
);
1412 add_zone_page_state(page_zone(page
),
1413 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1415 __SetPageSlab(page
);
1416 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1417 if (sk_memalloc_socks() && page_is_pfmemalloc(page
))
1418 SetPageSlabPfmemalloc(page
);
1420 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1421 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1424 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1426 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1433 * Interface to system's page release.
1435 static void kmem_freepages(struct kmem_cache
*cachep
, struct page
*page
)
1437 int order
= cachep
->gfporder
;
1438 unsigned long nr_freed
= (1 << order
);
1440 kmemcheck_free_shadow(page
, order
);
1442 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1443 sub_zone_page_state(page_zone(page
),
1444 NR_SLAB_RECLAIMABLE
, nr_freed
);
1446 sub_zone_page_state(page_zone(page
),
1447 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1449 BUG_ON(!PageSlab(page
));
1450 __ClearPageSlabPfmemalloc(page
);
1451 __ClearPageSlab(page
);
1452 page_mapcount_reset(page
);
1453 page
->mapping
= NULL
;
1455 if (current
->reclaim_state
)
1456 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1457 memcg_uncharge_slab(page
, order
, cachep
);
1458 __free_pages(page
, order
);
1461 static void kmem_rcu_free(struct rcu_head
*head
)
1463 struct kmem_cache
*cachep
;
1466 page
= container_of(head
, struct page
, rcu_head
);
1467 cachep
= page
->slab_cache
;
1469 kmem_freepages(cachep
, page
);
1473 static bool is_debug_pagealloc_cache(struct kmem_cache
*cachep
)
1475 if (debug_pagealloc_enabled() && OFF_SLAB(cachep
) &&
1476 (cachep
->size
% PAGE_SIZE
) == 0)
1482 #ifdef CONFIG_DEBUG_PAGEALLOC
1483 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1484 unsigned long caller
)
1486 int size
= cachep
->object_size
;
1488 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1490 if (size
< 5 * sizeof(unsigned long))
1493 *addr
++ = 0x12345678;
1495 *addr
++ = smp_processor_id();
1496 size
-= 3 * sizeof(unsigned long);
1498 unsigned long *sptr
= &caller
;
1499 unsigned long svalue
;
1501 while (!kstack_end(sptr
)) {
1503 if (kernel_text_address(svalue
)) {
1505 size
-= sizeof(unsigned long);
1506 if (size
<= sizeof(unsigned long))
1512 *addr
++ = 0x87654321;
1515 static void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1516 int map
, unsigned long caller
)
1518 if (!is_debug_pagealloc_cache(cachep
))
1522 store_stackinfo(cachep
, objp
, caller
);
1524 kernel_map_pages(virt_to_page(objp
), cachep
->size
/ PAGE_SIZE
, map
);
1528 static inline void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1529 int map
, unsigned long caller
) {}
1533 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1535 int size
= cachep
->object_size
;
1536 addr
= &((char *)addr
)[obj_offset(cachep
)];
1538 memset(addr
, val
, size
);
1539 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1542 static void dump_line(char *data
, int offset
, int limit
)
1545 unsigned char error
= 0;
1548 pr_err("%03x: ", offset
);
1549 for (i
= 0; i
< limit
; i
++) {
1550 if (data
[offset
+ i
] != POISON_FREE
) {
1551 error
= data
[offset
+ i
];
1555 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1556 &data
[offset
], limit
, 1);
1558 if (bad_count
== 1) {
1559 error
^= POISON_FREE
;
1560 if (!(error
& (error
- 1))) {
1561 pr_err("Single bit error detected. Probably bad RAM.\n");
1563 pr_err("Run memtest86+ or a similar memory test tool.\n");
1565 pr_err("Run a memory test tool.\n");
1574 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1579 if (cachep
->flags
& SLAB_RED_ZONE
) {
1580 pr_err("Redzone: 0x%llx/0x%llx\n",
1581 *dbg_redzone1(cachep
, objp
),
1582 *dbg_redzone2(cachep
, objp
));
1585 if (cachep
->flags
& SLAB_STORE_USER
) {
1586 pr_err("Last user: [<%p>](%pSR)\n",
1587 *dbg_userword(cachep
, objp
),
1588 *dbg_userword(cachep
, objp
));
1590 realobj
= (char *)objp
+ obj_offset(cachep
);
1591 size
= cachep
->object_size
;
1592 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1595 if (i
+ limit
> size
)
1597 dump_line(realobj
, i
, limit
);
1601 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1607 if (is_debug_pagealloc_cache(cachep
))
1610 realobj
= (char *)objp
+ obj_offset(cachep
);
1611 size
= cachep
->object_size
;
1613 for (i
= 0; i
< size
; i
++) {
1614 char exp
= POISON_FREE
;
1617 if (realobj
[i
] != exp
) {
1622 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1623 print_tainted(), cachep
->name
,
1625 print_objinfo(cachep
, objp
, 0);
1627 /* Hexdump the affected line */
1630 if (i
+ limit
> size
)
1632 dump_line(realobj
, i
, limit
);
1635 /* Limit to 5 lines */
1641 /* Print some data about the neighboring objects, if they
1644 struct page
*page
= virt_to_head_page(objp
);
1647 objnr
= obj_to_index(cachep
, page
, objp
);
1649 objp
= index_to_obj(cachep
, page
, objnr
- 1);
1650 realobj
= (char *)objp
+ obj_offset(cachep
);
1651 pr_err("Prev obj: start=%p, len=%d\n", realobj
, size
);
1652 print_objinfo(cachep
, objp
, 2);
1654 if (objnr
+ 1 < cachep
->num
) {
1655 objp
= index_to_obj(cachep
, page
, objnr
+ 1);
1656 realobj
= (char *)objp
+ obj_offset(cachep
);
1657 pr_err("Next obj: start=%p, len=%d\n", realobj
, size
);
1658 print_objinfo(cachep
, objp
, 2);
1665 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1670 if (OBJFREELIST_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
) {
1671 poison_obj(cachep
, page
->freelist
- obj_offset(cachep
),
1675 for (i
= 0; i
< cachep
->num
; i
++) {
1676 void *objp
= index_to_obj(cachep
, page
, i
);
1678 if (cachep
->flags
& SLAB_POISON
) {
1679 check_poison_obj(cachep
, objp
);
1680 slab_kernel_map(cachep
, objp
, 1, 0);
1682 if (cachep
->flags
& SLAB_RED_ZONE
) {
1683 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1684 slab_error(cachep
, "start of a freed object was overwritten");
1685 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1686 slab_error(cachep
, "end of a freed object was overwritten");
1691 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1698 * slab_destroy - destroy and release all objects in a slab
1699 * @cachep: cache pointer being destroyed
1700 * @page: page pointer being destroyed
1702 * Destroy all the objs in a slab page, and release the mem back to the system.
1703 * Before calling the slab page must have been unlinked from the cache. The
1704 * kmem_cache_node ->list_lock is not held/needed.
1706 static void slab_destroy(struct kmem_cache
*cachep
, struct page
*page
)
1710 freelist
= page
->freelist
;
1711 slab_destroy_debugcheck(cachep
, page
);
1712 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
1713 call_rcu(&page
->rcu_head
, kmem_rcu_free
);
1715 kmem_freepages(cachep
, page
);
1718 * From now on, we don't use freelist
1719 * although actual page can be freed in rcu context
1721 if (OFF_SLAB(cachep
))
1722 kmem_cache_free(cachep
->freelist_cache
, freelist
);
1725 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
)
1727 struct page
*page
, *n
;
1729 list_for_each_entry_safe(page
, n
, list
, lru
) {
1730 list_del(&page
->lru
);
1731 slab_destroy(cachep
, page
);
1736 * calculate_slab_order - calculate size (page order) of slabs
1737 * @cachep: pointer to the cache that is being created
1738 * @size: size of objects to be created in this cache.
1739 * @flags: slab allocation flags
1741 * Also calculates the number of objects per slab.
1743 * This could be made much more intelligent. For now, try to avoid using
1744 * high order pages for slabs. When the gfp() functions are more friendly
1745 * towards high-order requests, this should be changed.
1747 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1748 size_t size
, unsigned long flags
)
1750 size_t left_over
= 0;
1753 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
1757 num
= cache_estimate(gfporder
, size
, flags
, &remainder
);
1761 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1762 if (num
> SLAB_OBJ_MAX_NUM
)
1765 if (flags
& CFLGS_OFF_SLAB
) {
1766 struct kmem_cache
*freelist_cache
;
1767 size_t freelist_size
;
1769 freelist_size
= num
* sizeof(freelist_idx_t
);
1770 freelist_cache
= kmalloc_slab(freelist_size
, 0u);
1771 if (!freelist_cache
)
1775 * Needed to avoid possible looping condition
1778 if (OFF_SLAB(freelist_cache
))
1781 /* check if off slab has enough benefit */
1782 if (freelist_cache
->size
> cachep
->size
/ 2)
1786 /* Found something acceptable - save it away */
1788 cachep
->gfporder
= gfporder
;
1789 left_over
= remainder
;
1792 * A VFS-reclaimable slab tends to have most allocations
1793 * as GFP_NOFS and we really don't want to have to be allocating
1794 * higher-order pages when we are unable to shrink dcache.
1796 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1800 * Large number of objects is good, but very large slabs are
1801 * currently bad for the gfp()s.
1803 if (gfporder
>= slab_max_order
)
1807 * Acceptable internal fragmentation?
1809 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
1815 static struct array_cache __percpu
*alloc_kmem_cache_cpus(
1816 struct kmem_cache
*cachep
, int entries
, int batchcount
)
1820 struct array_cache __percpu
*cpu_cache
;
1822 size
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
1823 cpu_cache
= __alloc_percpu(size
, sizeof(void *));
1828 for_each_possible_cpu(cpu
) {
1829 init_arraycache(per_cpu_ptr(cpu_cache
, cpu
),
1830 entries
, batchcount
);
1836 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
1838 if (slab_state
>= FULL
)
1839 return enable_cpucache(cachep
, gfp
);
1841 cachep
->cpu_cache
= alloc_kmem_cache_cpus(cachep
, 1, 1);
1842 if (!cachep
->cpu_cache
)
1845 if (slab_state
== DOWN
) {
1846 /* Creation of first cache (kmem_cache). */
1847 set_up_node(kmem_cache
, CACHE_CACHE
);
1848 } else if (slab_state
== PARTIAL
) {
1849 /* For kmem_cache_node */
1850 set_up_node(cachep
, SIZE_NODE
);
1854 for_each_online_node(node
) {
1855 cachep
->node
[node
] = kmalloc_node(
1856 sizeof(struct kmem_cache_node
), gfp
, node
);
1857 BUG_ON(!cachep
->node
[node
]);
1858 kmem_cache_node_init(cachep
->node
[node
]);
1862 cachep
->node
[numa_mem_id()]->next_reap
=
1863 jiffies
+ REAPTIMEOUT_NODE
+
1864 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1866 cpu_cache_get(cachep
)->avail
= 0;
1867 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1868 cpu_cache_get(cachep
)->batchcount
= 1;
1869 cpu_cache_get(cachep
)->touched
= 0;
1870 cachep
->batchcount
= 1;
1871 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1875 unsigned long kmem_cache_flags(unsigned long object_size
,
1876 unsigned long flags
, const char *name
,
1877 void (*ctor
)(void *))
1883 __kmem_cache_alias(const char *name
, size_t size
, size_t align
,
1884 unsigned long flags
, void (*ctor
)(void *))
1886 struct kmem_cache
*cachep
;
1888 cachep
= find_mergeable(size
, align
, flags
, name
, ctor
);
1893 * Adjust the object sizes so that we clear
1894 * the complete object on kzalloc.
1896 cachep
->object_size
= max_t(int, cachep
->object_size
, size
);
1901 static bool set_objfreelist_slab_cache(struct kmem_cache
*cachep
,
1902 size_t size
, unsigned long flags
)
1908 if (cachep
->ctor
|| flags
& SLAB_DESTROY_BY_RCU
)
1911 left
= calculate_slab_order(cachep
, size
,
1912 flags
| CFLGS_OBJFREELIST_SLAB
);
1916 if (cachep
->num
* sizeof(freelist_idx_t
) > cachep
->object_size
)
1919 cachep
->colour
= left
/ cachep
->colour_off
;
1924 static bool set_off_slab_cache(struct kmem_cache
*cachep
,
1925 size_t size
, unsigned long flags
)
1932 * Always use on-slab management when SLAB_NOLEAKTRACE
1933 * to avoid recursive calls into kmemleak.
1935 if (flags
& SLAB_NOLEAKTRACE
)
1939 * Size is large, assume best to place the slab management obj
1940 * off-slab (should allow better packing of objs).
1942 left
= calculate_slab_order(cachep
, size
, flags
| CFLGS_OFF_SLAB
);
1947 * If the slab has been placed off-slab, and we have enough space then
1948 * move it on-slab. This is at the expense of any extra colouring.
1950 if (left
>= cachep
->num
* sizeof(freelist_idx_t
))
1953 cachep
->colour
= left
/ cachep
->colour_off
;
1958 static bool set_on_slab_cache(struct kmem_cache
*cachep
,
1959 size_t size
, unsigned long flags
)
1965 left
= calculate_slab_order(cachep
, size
, flags
);
1969 cachep
->colour
= left
/ cachep
->colour_off
;
1975 * __kmem_cache_create - Create a cache.
1976 * @cachep: cache management descriptor
1977 * @flags: SLAB flags
1979 * Returns a ptr to the cache on success, NULL on failure.
1980 * Cannot be called within a int, but can be interrupted.
1981 * The @ctor is run when new pages are allocated by the cache.
1985 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1986 * to catch references to uninitialised memory.
1988 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1989 * for buffer overruns.
1991 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1992 * cacheline. This can be beneficial if you're counting cycles as closely
1996 __kmem_cache_create (struct kmem_cache
*cachep
, unsigned long flags
)
1998 size_t ralign
= BYTES_PER_WORD
;
2001 size_t size
= cachep
->size
;
2006 * Enable redzoning and last user accounting, except for caches with
2007 * large objects, if the increased size would increase the object size
2008 * above the next power of two: caches with object sizes just above a
2009 * power of two have a significant amount of internal fragmentation.
2011 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2012 2 * sizeof(unsigned long long)))
2013 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2014 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2015 flags
|= SLAB_POISON
;
2020 * Check that size is in terms of words. This is needed to avoid
2021 * unaligned accesses for some archs when redzoning is used, and makes
2022 * sure any on-slab bufctl's are also correctly aligned.
2024 if (size
& (BYTES_PER_WORD
- 1)) {
2025 size
+= (BYTES_PER_WORD
- 1);
2026 size
&= ~(BYTES_PER_WORD
- 1);
2029 if (flags
& SLAB_RED_ZONE
) {
2030 ralign
= REDZONE_ALIGN
;
2031 /* If redzoning, ensure that the second redzone is suitably
2032 * aligned, by adjusting the object size accordingly. */
2033 size
+= REDZONE_ALIGN
- 1;
2034 size
&= ~(REDZONE_ALIGN
- 1);
2037 /* 3) caller mandated alignment */
2038 if (ralign
< cachep
->align
) {
2039 ralign
= cachep
->align
;
2041 /* disable debug if necessary */
2042 if (ralign
> __alignof__(unsigned long long))
2043 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2047 cachep
->align
= ralign
;
2048 cachep
->colour_off
= cache_line_size();
2049 /* Offset must be a multiple of the alignment. */
2050 if (cachep
->colour_off
< cachep
->align
)
2051 cachep
->colour_off
= cachep
->align
;
2053 if (slab_is_available())
2061 * Both debugging options require word-alignment which is calculated
2064 if (flags
& SLAB_RED_ZONE
) {
2065 /* add space for red zone words */
2066 cachep
->obj_offset
+= sizeof(unsigned long long);
2067 size
+= 2 * sizeof(unsigned long long);
2069 if (flags
& SLAB_STORE_USER
) {
2070 /* user store requires one word storage behind the end of
2071 * the real object. But if the second red zone needs to be
2072 * aligned to 64 bits, we must allow that much space.
2074 if (flags
& SLAB_RED_ZONE
)
2075 size
+= REDZONE_ALIGN
;
2077 size
+= BYTES_PER_WORD
;
2081 kasan_cache_create(cachep
, &size
, &flags
);
2083 size
= ALIGN(size
, cachep
->align
);
2085 * We should restrict the number of objects in a slab to implement
2086 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2088 if (FREELIST_BYTE_INDEX
&& size
< SLAB_OBJ_MIN_SIZE
)
2089 size
= ALIGN(SLAB_OBJ_MIN_SIZE
, cachep
->align
);
2093 * To activate debug pagealloc, off-slab management is necessary
2094 * requirement. In early phase of initialization, small sized slab
2095 * doesn't get initialized so it would not be possible. So, we need
2096 * to check size >= 256. It guarantees that all necessary small
2097 * sized slab is initialized in current slab initialization sequence.
2099 if (debug_pagealloc_enabled() && (flags
& SLAB_POISON
) &&
2100 size
>= 256 && cachep
->object_size
> cache_line_size()) {
2101 if (size
< PAGE_SIZE
|| size
% PAGE_SIZE
== 0) {
2102 size_t tmp_size
= ALIGN(size
, PAGE_SIZE
);
2104 if (set_off_slab_cache(cachep
, tmp_size
, flags
)) {
2105 flags
|= CFLGS_OFF_SLAB
;
2106 cachep
->obj_offset
+= tmp_size
- size
;
2114 if (set_objfreelist_slab_cache(cachep
, size
, flags
)) {
2115 flags
|= CFLGS_OBJFREELIST_SLAB
;
2119 if (set_off_slab_cache(cachep
, size
, flags
)) {
2120 flags
|= CFLGS_OFF_SLAB
;
2124 if (set_on_slab_cache(cachep
, size
, flags
))
2130 cachep
->freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
2131 cachep
->flags
= flags
;
2132 cachep
->allocflags
= __GFP_COMP
;
2133 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2134 cachep
->allocflags
|= GFP_DMA
;
2135 cachep
->size
= size
;
2136 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2140 * If we're going to use the generic kernel_map_pages()
2141 * poisoning, then it's going to smash the contents of
2142 * the redzone and userword anyhow, so switch them off.
2144 if (IS_ENABLED(CONFIG_PAGE_POISONING
) &&
2145 (cachep
->flags
& SLAB_POISON
) &&
2146 is_debug_pagealloc_cache(cachep
))
2147 cachep
->flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2150 if (OFF_SLAB(cachep
)) {
2151 cachep
->freelist_cache
=
2152 kmalloc_slab(cachep
->freelist_size
, 0u);
2155 err
= setup_cpu_cache(cachep
, gfp
);
2157 __kmem_cache_release(cachep
);
2165 static void check_irq_off(void)
2167 BUG_ON(!irqs_disabled());
2170 static void check_irq_on(void)
2172 BUG_ON(irqs_disabled());
2175 static void check_mutex_acquired(void)
2177 BUG_ON(!mutex_is_locked(&slab_mutex
));
2180 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2184 assert_spin_locked(&get_node(cachep
, numa_mem_id())->list_lock
);
2188 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2192 assert_spin_locked(&get_node(cachep
, node
)->list_lock
);
2197 #define check_irq_off() do { } while(0)
2198 #define check_irq_on() do { } while(0)
2199 #define check_mutex_acquired() do { } while(0)
2200 #define check_spinlock_acquired(x) do { } while(0)
2201 #define check_spinlock_acquired_node(x, y) do { } while(0)
2204 static void drain_array_locked(struct kmem_cache
*cachep
, struct array_cache
*ac
,
2205 int node
, bool free_all
, struct list_head
*list
)
2209 if (!ac
|| !ac
->avail
)
2212 tofree
= free_all
? ac
->avail
: (ac
->limit
+ 4) / 5;
2213 if (tofree
> ac
->avail
)
2214 tofree
= (ac
->avail
+ 1) / 2;
2216 free_block(cachep
, ac
->entry
, tofree
, node
, list
);
2217 ac
->avail
-= tofree
;
2218 memmove(ac
->entry
, &(ac
->entry
[tofree
]), sizeof(void *) * ac
->avail
);
2221 static void do_drain(void *arg
)
2223 struct kmem_cache
*cachep
= arg
;
2224 struct array_cache
*ac
;
2225 int node
= numa_mem_id();
2226 struct kmem_cache_node
*n
;
2230 ac
= cpu_cache_get(cachep
);
2231 n
= get_node(cachep
, node
);
2232 spin_lock(&n
->list_lock
);
2233 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
2234 spin_unlock(&n
->list_lock
);
2235 slabs_destroy(cachep
, &list
);
2239 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2241 struct kmem_cache_node
*n
;
2245 on_each_cpu(do_drain
, cachep
, 1);
2247 for_each_kmem_cache_node(cachep
, node
, n
)
2249 drain_alien_cache(cachep
, n
->alien
);
2251 for_each_kmem_cache_node(cachep
, node
, n
) {
2252 spin_lock_irq(&n
->list_lock
);
2253 drain_array_locked(cachep
, n
->shared
, node
, true, &list
);
2254 spin_unlock_irq(&n
->list_lock
);
2256 slabs_destroy(cachep
, &list
);
2261 * Remove slabs from the list of free slabs.
2262 * Specify the number of slabs to drain in tofree.
2264 * Returns the actual number of slabs released.
2266 static int drain_freelist(struct kmem_cache
*cache
,
2267 struct kmem_cache_node
*n
, int tofree
)
2269 struct list_head
*p
;
2274 while (nr_freed
< tofree
&& !list_empty(&n
->slabs_free
)) {
2276 spin_lock_irq(&n
->list_lock
);
2277 p
= n
->slabs_free
.prev
;
2278 if (p
== &n
->slabs_free
) {
2279 spin_unlock_irq(&n
->list_lock
);
2283 page
= list_entry(p
, struct page
, lru
);
2284 list_del(&page
->lru
);
2286 * Safe to drop the lock. The slab is no longer linked
2289 n
->free_objects
-= cache
->num
;
2290 spin_unlock_irq(&n
->list_lock
);
2291 slab_destroy(cache
, page
);
2298 int __kmem_cache_shrink(struct kmem_cache
*cachep
, bool deactivate
)
2302 struct kmem_cache_node
*n
;
2304 drain_cpu_caches(cachep
);
2307 for_each_kmem_cache_node(cachep
, node
, n
) {
2308 drain_freelist(cachep
, n
, INT_MAX
);
2310 ret
+= !list_empty(&n
->slabs_full
) ||
2311 !list_empty(&n
->slabs_partial
);
2313 return (ret
? 1 : 0);
2316 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2318 return __kmem_cache_shrink(cachep
, false);
2321 void __kmem_cache_release(struct kmem_cache
*cachep
)
2324 struct kmem_cache_node
*n
;
2326 free_percpu(cachep
->cpu_cache
);
2328 /* NUMA: free the node structures */
2329 for_each_kmem_cache_node(cachep
, i
, n
) {
2331 free_alien_cache(n
->alien
);
2333 cachep
->node
[i
] = NULL
;
2338 * Get the memory for a slab management obj.
2340 * For a slab cache when the slab descriptor is off-slab, the
2341 * slab descriptor can't come from the same cache which is being created,
2342 * Because if it is the case, that means we defer the creation of
2343 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2344 * And we eventually call down to __kmem_cache_create(), which
2345 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2346 * This is a "chicken-and-egg" problem.
2348 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2349 * which are all initialized during kmem_cache_init().
2351 static void *alloc_slabmgmt(struct kmem_cache
*cachep
,
2352 struct page
*page
, int colour_off
,
2353 gfp_t local_flags
, int nodeid
)
2356 void *addr
= page_address(page
);
2358 page
->s_mem
= addr
+ colour_off
;
2361 if (OBJFREELIST_SLAB(cachep
))
2363 else if (OFF_SLAB(cachep
)) {
2364 /* Slab management obj is off-slab. */
2365 freelist
= kmem_cache_alloc_node(cachep
->freelist_cache
,
2366 local_flags
, nodeid
);
2370 /* We will use last bytes at the slab for freelist */
2371 freelist
= addr
+ (PAGE_SIZE
<< cachep
->gfporder
) -
2372 cachep
->freelist_size
;
2378 static inline freelist_idx_t
get_free_obj(struct page
*page
, unsigned int idx
)
2380 return ((freelist_idx_t
*)page
->freelist
)[idx
];
2383 static inline void set_free_obj(struct page
*page
,
2384 unsigned int idx
, freelist_idx_t val
)
2386 ((freelist_idx_t
*)(page
->freelist
))[idx
] = val
;
2389 static void cache_init_objs_debug(struct kmem_cache
*cachep
, struct page
*page
)
2394 for (i
= 0; i
< cachep
->num
; i
++) {
2395 void *objp
= index_to_obj(cachep
, page
, i
);
2397 if (cachep
->flags
& SLAB_STORE_USER
)
2398 *dbg_userword(cachep
, objp
) = NULL
;
2400 if (cachep
->flags
& SLAB_RED_ZONE
) {
2401 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2402 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2405 * Constructors are not allowed to allocate memory from the same
2406 * cache which they are a constructor for. Otherwise, deadlock.
2407 * They must also be threaded.
2409 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
)) {
2410 kasan_unpoison_object_data(cachep
,
2411 objp
+ obj_offset(cachep
));
2412 cachep
->ctor(objp
+ obj_offset(cachep
));
2413 kasan_poison_object_data(
2414 cachep
, objp
+ obj_offset(cachep
));
2417 if (cachep
->flags
& SLAB_RED_ZONE
) {
2418 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2419 slab_error(cachep
, "constructor overwrote the end of an object");
2420 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2421 slab_error(cachep
, "constructor overwrote the start of an object");
2423 /* need to poison the objs? */
2424 if (cachep
->flags
& SLAB_POISON
) {
2425 poison_obj(cachep
, objp
, POISON_FREE
);
2426 slab_kernel_map(cachep
, objp
, 0, 0);
2432 static void cache_init_objs(struct kmem_cache
*cachep
,
2438 cache_init_objs_debug(cachep
, page
);
2440 if (OBJFREELIST_SLAB(cachep
)) {
2441 page
->freelist
= index_to_obj(cachep
, page
, cachep
->num
- 1) +
2445 for (i
= 0; i
< cachep
->num
; i
++) {
2446 /* constructor could break poison info */
2447 if (DEBUG
== 0 && cachep
->ctor
) {
2448 objp
= index_to_obj(cachep
, page
, i
);
2449 kasan_unpoison_object_data(cachep
, objp
);
2451 kasan_poison_object_data(cachep
, objp
);
2454 set_free_obj(page
, i
, i
);
2458 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2460 if (CONFIG_ZONE_DMA_FLAG
) {
2461 if (flags
& GFP_DMA
)
2462 BUG_ON(!(cachep
->allocflags
& GFP_DMA
));
2464 BUG_ON(cachep
->allocflags
& GFP_DMA
);
2468 static void *slab_get_obj(struct kmem_cache
*cachep
, struct page
*page
)
2472 objp
= index_to_obj(cachep
, page
, get_free_obj(page
, page
->active
));
2476 if (cachep
->flags
& SLAB_STORE_USER
)
2477 set_store_user_dirty(cachep
);
2483 static void slab_put_obj(struct kmem_cache
*cachep
,
2484 struct page
*page
, void *objp
)
2486 unsigned int objnr
= obj_to_index(cachep
, page
, objp
);
2490 /* Verify double free bug */
2491 for (i
= page
->active
; i
< cachep
->num
; i
++) {
2492 if (get_free_obj(page
, i
) == objnr
) {
2493 pr_err("slab: double free detected in cache '%s', objp %p\n",
2494 cachep
->name
, objp
);
2500 if (!page
->freelist
)
2501 page
->freelist
= objp
+ obj_offset(cachep
);
2503 set_free_obj(page
, page
->active
, objnr
);
2507 * Map pages beginning at addr to the given cache and slab. This is required
2508 * for the slab allocator to be able to lookup the cache and slab of a
2509 * virtual address for kfree, ksize, and slab debugging.
2511 static void slab_map_pages(struct kmem_cache
*cache
, struct page
*page
,
2514 page
->slab_cache
= cache
;
2515 page
->freelist
= freelist
;
2519 * Grow (by 1) the number of slabs within a cache. This is called by
2520 * kmem_cache_alloc() when there are no active objs left in a cache.
2522 static int cache_grow(struct kmem_cache
*cachep
,
2523 gfp_t flags
, int nodeid
, struct page
*page
)
2528 struct kmem_cache_node
*n
;
2531 * Be lazy and only check for valid flags here, keeping it out of the
2532 * critical path in kmem_cache_alloc().
2534 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
2535 pr_emerg("gfp: %u\n", flags
& GFP_SLAB_BUG_MASK
);
2538 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2540 /* Take the node list lock to change the colour_next on this node */
2542 n
= get_node(cachep
, nodeid
);
2543 spin_lock(&n
->list_lock
);
2545 /* Get colour for the slab, and cal the next value. */
2546 offset
= n
->colour_next
;
2548 if (n
->colour_next
>= cachep
->colour
)
2550 spin_unlock(&n
->list_lock
);
2552 offset
*= cachep
->colour_off
;
2554 if (gfpflags_allow_blocking(local_flags
))
2558 * The test for missing atomic flag is performed here, rather than
2559 * the more obvious place, simply to reduce the critical path length
2560 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2561 * will eventually be caught here (where it matters).
2563 kmem_flagcheck(cachep
, flags
);
2566 * Get mem for the objs. Attempt to allocate a physical page from
2570 page
= kmem_getpages(cachep
, local_flags
, nodeid
);
2574 /* Get slab management. */
2575 freelist
= alloc_slabmgmt(cachep
, page
, offset
,
2576 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
2577 if (OFF_SLAB(cachep
) && !freelist
)
2580 slab_map_pages(cachep
, page
, freelist
);
2582 kasan_poison_slab(page
);
2583 cache_init_objs(cachep
, page
);
2585 if (gfpflags_allow_blocking(local_flags
))
2586 local_irq_disable();
2588 spin_lock(&n
->list_lock
);
2590 /* Make slab active. */
2591 list_add_tail(&page
->lru
, &(n
->slabs_free
));
2592 STATS_INC_GROWN(cachep
);
2593 n
->free_objects
+= cachep
->num
;
2594 spin_unlock(&n
->list_lock
);
2597 kmem_freepages(cachep
, page
);
2599 if (gfpflags_allow_blocking(local_flags
))
2600 local_irq_disable();
2607 * Perform extra freeing checks:
2608 * - detect bad pointers.
2609 * - POISON/RED_ZONE checking
2611 static void kfree_debugcheck(const void *objp
)
2613 if (!virt_addr_valid(objp
)) {
2614 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2615 (unsigned long)objp
);
2620 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2622 unsigned long long redzone1
, redzone2
;
2624 redzone1
= *dbg_redzone1(cache
, obj
);
2625 redzone2
= *dbg_redzone2(cache
, obj
);
2630 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2633 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2634 slab_error(cache
, "double free detected");
2636 slab_error(cache
, "memory outside object was overwritten");
2638 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2639 obj
, redzone1
, redzone2
);
2642 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2643 unsigned long caller
)
2648 BUG_ON(virt_to_cache(objp
) != cachep
);
2650 objp
-= obj_offset(cachep
);
2651 kfree_debugcheck(objp
);
2652 page
= virt_to_head_page(objp
);
2654 if (cachep
->flags
& SLAB_RED_ZONE
) {
2655 verify_redzone_free(cachep
, objp
);
2656 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2657 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2659 if (cachep
->flags
& SLAB_STORE_USER
) {
2660 set_store_user_dirty(cachep
);
2661 *dbg_userword(cachep
, objp
) = (void *)caller
;
2664 objnr
= obj_to_index(cachep
, page
, objp
);
2666 BUG_ON(objnr
>= cachep
->num
);
2667 BUG_ON(objp
!= index_to_obj(cachep
, page
, objnr
));
2669 if (cachep
->flags
& SLAB_POISON
) {
2670 poison_obj(cachep
, objp
, POISON_FREE
);
2671 slab_kernel_map(cachep
, objp
, 0, caller
);
2677 #define kfree_debugcheck(x) do { } while(0)
2678 #define cache_free_debugcheck(x,objp,z) (objp)
2681 static inline void fixup_objfreelist_debug(struct kmem_cache
*cachep
,
2689 objp
= next
- obj_offset(cachep
);
2690 next
= *(void **)next
;
2691 poison_obj(cachep
, objp
, POISON_FREE
);
2696 static inline void fixup_slab_list(struct kmem_cache
*cachep
,
2697 struct kmem_cache_node
*n
, struct page
*page
,
2700 /* move slabp to correct slabp list: */
2701 list_del(&page
->lru
);
2702 if (page
->active
== cachep
->num
) {
2703 list_add(&page
->lru
, &n
->slabs_full
);
2704 if (OBJFREELIST_SLAB(cachep
)) {
2706 /* Poisoning will be done without holding the lock */
2707 if (cachep
->flags
& SLAB_POISON
) {
2708 void **objp
= page
->freelist
;
2714 page
->freelist
= NULL
;
2717 list_add(&page
->lru
, &n
->slabs_partial
);
2720 /* Try to find non-pfmemalloc slab if needed */
2721 static noinline
struct page
*get_valid_first_slab(struct kmem_cache_node
*n
,
2722 struct page
*page
, bool pfmemalloc
)
2730 if (!PageSlabPfmemalloc(page
))
2733 /* No need to keep pfmemalloc slab if we have enough free objects */
2734 if (n
->free_objects
> n
->free_limit
) {
2735 ClearPageSlabPfmemalloc(page
);
2739 /* Move pfmemalloc slab to the end of list to speed up next search */
2740 list_del(&page
->lru
);
2742 list_add_tail(&page
->lru
, &n
->slabs_free
);
2744 list_add_tail(&page
->lru
, &n
->slabs_partial
);
2746 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
2747 if (!PageSlabPfmemalloc(page
))
2751 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
2752 if (!PageSlabPfmemalloc(page
))
2759 static struct page
*get_first_slab(struct kmem_cache_node
*n
, bool pfmemalloc
)
2763 page
= list_first_entry_or_null(&n
->slabs_partial
,
2766 n
->free_touched
= 1;
2767 page
= list_first_entry_or_null(&n
->slabs_free
,
2771 if (sk_memalloc_socks())
2772 return get_valid_first_slab(n
, page
, pfmemalloc
);
2777 static noinline
void *cache_alloc_pfmemalloc(struct kmem_cache
*cachep
,
2778 struct kmem_cache_node
*n
, gfp_t flags
)
2784 if (!gfp_pfmemalloc_allowed(flags
))
2787 spin_lock(&n
->list_lock
);
2788 page
= get_first_slab(n
, true);
2790 spin_unlock(&n
->list_lock
);
2794 obj
= slab_get_obj(cachep
, page
);
2797 fixup_slab_list(cachep
, n
, page
, &list
);
2799 spin_unlock(&n
->list_lock
);
2800 fixup_objfreelist_debug(cachep
, &list
);
2805 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2808 struct kmem_cache_node
*n
;
2809 struct array_cache
*ac
;
2814 node
= numa_mem_id();
2817 ac
= cpu_cache_get(cachep
);
2818 batchcount
= ac
->batchcount
;
2819 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2821 * If there was little recent activity on this cache, then
2822 * perform only a partial refill. Otherwise we could generate
2825 batchcount
= BATCHREFILL_LIMIT
;
2827 n
= get_node(cachep
, node
);
2829 BUG_ON(ac
->avail
> 0 || !n
);
2830 spin_lock(&n
->list_lock
);
2832 /* See if we can refill from the shared array */
2833 if (n
->shared
&& transfer_objects(ac
, n
->shared
, batchcount
)) {
2834 n
->shared
->touched
= 1;
2838 while (batchcount
> 0) {
2840 /* Get slab alloc is to come from. */
2841 page
= get_first_slab(n
, false);
2845 check_spinlock_acquired(cachep
);
2848 * The slab was either on partial or free list so
2849 * there must be at least one object available for
2852 BUG_ON(page
->active
>= cachep
->num
);
2854 while (page
->active
< cachep
->num
&& batchcount
--) {
2855 STATS_INC_ALLOCED(cachep
);
2856 STATS_INC_ACTIVE(cachep
);
2857 STATS_SET_HIGH(cachep
);
2859 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, page
);
2862 fixup_slab_list(cachep
, n
, page
, &list
);
2866 n
->free_objects
-= ac
->avail
;
2868 spin_unlock(&n
->list_lock
);
2869 fixup_objfreelist_debug(cachep
, &list
);
2871 if (unlikely(!ac
->avail
)) {
2874 /* Check if we can use obj in pfmemalloc slab */
2875 if (sk_memalloc_socks()) {
2876 void *obj
= cache_alloc_pfmemalloc(cachep
, n
, flags
);
2882 x
= cache_grow(cachep
, gfp_exact_node(flags
), node
, NULL
);
2884 /* cache_grow can reenable interrupts, then ac could change. */
2885 ac
= cpu_cache_get(cachep
);
2886 node
= numa_mem_id();
2888 /* no objects in sight? abort */
2889 if (!x
&& ac
->avail
== 0)
2892 if (!ac
->avail
) /* objects refilled by interrupt? */
2897 return ac
->entry
[--ac
->avail
];
2900 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2903 might_sleep_if(gfpflags_allow_blocking(flags
));
2905 kmem_flagcheck(cachep
, flags
);
2910 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2911 gfp_t flags
, void *objp
, unsigned long caller
)
2915 if (cachep
->flags
& SLAB_POISON
) {
2916 check_poison_obj(cachep
, objp
);
2917 slab_kernel_map(cachep
, objp
, 1, 0);
2918 poison_obj(cachep
, objp
, POISON_INUSE
);
2920 if (cachep
->flags
& SLAB_STORE_USER
)
2921 *dbg_userword(cachep
, objp
) = (void *)caller
;
2923 if (cachep
->flags
& SLAB_RED_ZONE
) {
2924 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
2925 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2926 slab_error(cachep
, "double free, or memory outside object was overwritten");
2927 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2928 objp
, *dbg_redzone1(cachep
, objp
),
2929 *dbg_redzone2(cachep
, objp
));
2931 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
2932 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
2935 objp
+= obj_offset(cachep
);
2936 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
2938 if (ARCH_SLAB_MINALIGN
&&
2939 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
2940 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
2941 objp
, (int)ARCH_SLAB_MINALIGN
);
2946 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2949 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2952 struct array_cache
*ac
;
2956 ac
= cpu_cache_get(cachep
);
2957 if (likely(ac
->avail
)) {
2959 objp
= ac
->entry
[--ac
->avail
];
2961 STATS_INC_ALLOCHIT(cachep
);
2965 STATS_INC_ALLOCMISS(cachep
);
2966 objp
= cache_alloc_refill(cachep
, flags
);
2968 * the 'ac' may be updated by cache_alloc_refill(),
2969 * and kmemleak_erase() requires its correct value.
2971 ac
= cpu_cache_get(cachep
);
2975 * To avoid a false negative, if an object that is in one of the
2976 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2977 * treat the array pointers as a reference to the object.
2980 kmemleak_erase(&ac
->entry
[ac
->avail
]);
2986 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
2988 * If we are in_interrupt, then process context, including cpusets and
2989 * mempolicy, may not apply and should not be used for allocation policy.
2991 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2993 int nid_alloc
, nid_here
;
2995 if (in_interrupt() || (flags
& __GFP_THISNODE
))
2997 nid_alloc
= nid_here
= numa_mem_id();
2998 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
2999 nid_alloc
= cpuset_slab_spread_node();
3000 else if (current
->mempolicy
)
3001 nid_alloc
= mempolicy_slab_node();
3002 if (nid_alloc
!= nid_here
)
3003 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3008 * Fallback function if there was no memory available and no objects on a
3009 * certain node and fall back is permitted. First we scan all the
3010 * available node for available objects. If that fails then we
3011 * perform an allocation without specifying a node. This allows the page
3012 * allocator to do its reclaim / fallback magic. We then insert the
3013 * slab into the proper nodelist and then allocate from it.
3015 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3017 struct zonelist
*zonelist
;
3021 enum zone_type high_zoneidx
= gfp_zone(flags
);
3024 unsigned int cpuset_mems_cookie
;
3026 if (flags
& __GFP_THISNODE
)
3029 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3032 cpuset_mems_cookie
= read_mems_allowed_begin();
3033 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
3037 * Look through allowed nodes for objects available
3038 * from existing per node queues.
3040 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3041 nid
= zone_to_nid(zone
);
3043 if (cpuset_zone_allowed(zone
, flags
) &&
3044 get_node(cache
, nid
) &&
3045 get_node(cache
, nid
)->free_objects
) {
3046 obj
= ____cache_alloc_node(cache
,
3047 gfp_exact_node(flags
), nid
);
3055 * This allocation will be performed within the constraints
3056 * of the current cpuset / memory policy requirements.
3057 * We may trigger various forms of reclaim on the allowed
3058 * set and go into memory reserves if necessary.
3062 if (gfpflags_allow_blocking(local_flags
))
3064 kmem_flagcheck(cache
, flags
);
3065 page
= kmem_getpages(cache
, local_flags
, numa_mem_id());
3066 if (gfpflags_allow_blocking(local_flags
))
3067 local_irq_disable();
3070 * Insert into the appropriate per node queues
3072 nid
= page_to_nid(page
);
3073 if (cache_grow(cache
, flags
, nid
, page
)) {
3074 obj
= ____cache_alloc_node(cache
,
3075 gfp_exact_node(flags
), nid
);
3078 * Another processor may allocate the
3079 * objects in the slab since we are
3080 * not holding any locks.
3084 /* cache_grow already freed obj */
3090 if (unlikely(!obj
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3096 * A interface to enable slab creation on nodeid
3098 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3102 struct kmem_cache_node
*n
;
3107 VM_BUG_ON(nodeid
< 0 || nodeid
>= MAX_NUMNODES
);
3108 n
= get_node(cachep
, nodeid
);
3113 spin_lock(&n
->list_lock
);
3114 page
= get_first_slab(n
, false);
3118 check_spinlock_acquired_node(cachep
, nodeid
);
3120 STATS_INC_NODEALLOCS(cachep
);
3121 STATS_INC_ACTIVE(cachep
);
3122 STATS_SET_HIGH(cachep
);
3124 BUG_ON(page
->active
== cachep
->num
);
3126 obj
= slab_get_obj(cachep
, page
);
3129 fixup_slab_list(cachep
, n
, page
, &list
);
3131 spin_unlock(&n
->list_lock
);
3132 fixup_objfreelist_debug(cachep
, &list
);
3136 spin_unlock(&n
->list_lock
);
3137 x
= cache_grow(cachep
, gfp_exact_node(flags
), nodeid
, NULL
);
3141 return fallback_alloc(cachep
, flags
);
3147 static __always_inline
void *
3148 slab_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3149 unsigned long caller
)
3151 unsigned long save_flags
;
3153 int slab_node
= numa_mem_id();
3155 flags
&= gfp_allowed_mask
;
3156 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3157 if (unlikely(!cachep
))
3160 cache_alloc_debugcheck_before(cachep
, flags
);
3161 local_irq_save(save_flags
);
3163 if (nodeid
== NUMA_NO_NODE
)
3166 if (unlikely(!get_node(cachep
, nodeid
))) {
3167 /* Node not bootstrapped yet */
3168 ptr
= fallback_alloc(cachep
, flags
);
3172 if (nodeid
== slab_node
) {
3174 * Use the locally cached objects if possible.
3175 * However ____cache_alloc does not allow fallback
3176 * to other nodes. It may fail while we still have
3177 * objects on other nodes available.
3179 ptr
= ____cache_alloc(cachep
, flags
);
3183 /* ___cache_alloc_node can fall back to other nodes */
3184 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3186 local_irq_restore(save_flags
);
3187 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3189 if (unlikely(flags
& __GFP_ZERO
) && ptr
)
3190 memset(ptr
, 0, cachep
->object_size
);
3192 slab_post_alloc_hook(cachep
, flags
, 1, &ptr
);
3196 static __always_inline
void *
3197 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3201 if (current
->mempolicy
|| cpuset_do_slab_mem_spread()) {
3202 objp
= alternate_node_alloc(cache
, flags
);
3206 objp
= ____cache_alloc(cache
, flags
);
3209 * We may just have run out of memory on the local node.
3210 * ____cache_alloc_node() knows how to locate memory on other nodes
3213 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3220 static __always_inline
void *
3221 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3223 return ____cache_alloc(cachep
, flags
);
3226 #endif /* CONFIG_NUMA */
3228 static __always_inline
void *
3229 slab_alloc(struct kmem_cache
*cachep
, gfp_t flags
, unsigned long caller
)
3231 unsigned long save_flags
;
3234 flags
&= gfp_allowed_mask
;
3235 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3236 if (unlikely(!cachep
))
3239 cache_alloc_debugcheck_before(cachep
, flags
);
3240 local_irq_save(save_flags
);
3241 objp
= __do_cache_alloc(cachep
, flags
);
3242 local_irq_restore(save_flags
);
3243 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3246 if (unlikely(flags
& __GFP_ZERO
) && objp
)
3247 memset(objp
, 0, cachep
->object_size
);
3249 slab_post_alloc_hook(cachep
, flags
, 1, &objp
);
3254 * Caller needs to acquire correct kmem_cache_node's list_lock
3255 * @list: List of detached free slabs should be freed by caller
3257 static void free_block(struct kmem_cache
*cachep
, void **objpp
,
3258 int nr_objects
, int node
, struct list_head
*list
)
3261 struct kmem_cache_node
*n
= get_node(cachep
, node
);
3263 for (i
= 0; i
< nr_objects
; i
++) {
3269 page
= virt_to_head_page(objp
);
3270 list_del(&page
->lru
);
3271 check_spinlock_acquired_node(cachep
, node
);
3272 slab_put_obj(cachep
, page
, objp
);
3273 STATS_DEC_ACTIVE(cachep
);
3276 /* fixup slab chains */
3277 if (page
->active
== 0) {
3278 if (n
->free_objects
> n
->free_limit
) {
3279 n
->free_objects
-= cachep
->num
;
3280 list_add_tail(&page
->lru
, list
);
3282 list_add(&page
->lru
, &n
->slabs_free
);
3285 /* Unconditionally move a slab to the end of the
3286 * partial list on free - maximum time for the
3287 * other objects to be freed, too.
3289 list_add_tail(&page
->lru
, &n
->slabs_partial
);
3294 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3297 struct kmem_cache_node
*n
;
3298 int node
= numa_mem_id();
3301 batchcount
= ac
->batchcount
;
3304 n
= get_node(cachep
, node
);
3305 spin_lock(&n
->list_lock
);
3307 struct array_cache
*shared_array
= n
->shared
;
3308 int max
= shared_array
->limit
- shared_array
->avail
;
3310 if (batchcount
> max
)
3312 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3313 ac
->entry
, sizeof(void *) * batchcount
);
3314 shared_array
->avail
+= batchcount
;
3319 free_block(cachep
, ac
->entry
, batchcount
, node
, &list
);
3326 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
3327 BUG_ON(page
->active
);
3331 STATS_SET_FREEABLE(cachep
, i
);
3334 spin_unlock(&n
->list_lock
);
3335 slabs_destroy(cachep
, &list
);
3336 ac
->avail
-= batchcount
;
3337 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3341 * Release an obj back to its cache. If the obj has a constructed state, it must
3342 * be in this state _before_ it is released. Called with disabled ints.
3344 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3345 unsigned long caller
)
3347 struct array_cache
*ac
= cpu_cache_get(cachep
);
3349 kasan_slab_free(cachep
, objp
);
3352 kmemleak_free_recursive(objp
, cachep
->flags
);
3353 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3355 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3358 * Skip calling cache_free_alien() when the platform is not numa.
3359 * This will avoid cache misses that happen while accessing slabp (which
3360 * is per page memory reference) to get nodeid. Instead use a global
3361 * variable to skip the call, which is mostly likely to be present in
3364 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3367 if (ac
->avail
< ac
->limit
) {
3368 STATS_INC_FREEHIT(cachep
);
3370 STATS_INC_FREEMISS(cachep
);
3371 cache_flusharray(cachep
, ac
);
3374 if (sk_memalloc_socks()) {
3375 struct page
*page
= virt_to_head_page(objp
);
3377 if (unlikely(PageSlabPfmemalloc(page
))) {
3378 cache_free_pfmemalloc(cachep
, page
, objp
);
3383 ac
->entry
[ac
->avail
++] = objp
;
3387 * kmem_cache_alloc - Allocate an object
3388 * @cachep: The cache to allocate from.
3389 * @flags: See kmalloc().
3391 * Allocate an object from this cache. The flags are only relevant
3392 * if the cache has no available objects.
3394 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3396 void *ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3398 kasan_slab_alloc(cachep
, ret
, flags
);
3399 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3400 cachep
->object_size
, cachep
->size
, flags
);
3404 EXPORT_SYMBOL(kmem_cache_alloc
);
3406 static __always_inline
void
3407 cache_alloc_debugcheck_after_bulk(struct kmem_cache
*s
, gfp_t flags
,
3408 size_t size
, void **p
, unsigned long caller
)
3412 for (i
= 0; i
< size
; i
++)
3413 p
[i
] = cache_alloc_debugcheck_after(s
, flags
, p
[i
], caller
);
3416 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3421 s
= slab_pre_alloc_hook(s
, flags
);
3425 cache_alloc_debugcheck_before(s
, flags
);
3427 local_irq_disable();
3428 for (i
= 0; i
< size
; i
++) {
3429 void *objp
= __do_cache_alloc(s
, flags
);
3431 if (unlikely(!objp
))
3437 cache_alloc_debugcheck_after_bulk(s
, flags
, size
, p
, _RET_IP_
);
3439 /* Clear memory outside IRQ disabled section */
3440 if (unlikely(flags
& __GFP_ZERO
))
3441 for (i
= 0; i
< size
; i
++)
3442 memset(p
[i
], 0, s
->object_size
);
3444 slab_post_alloc_hook(s
, flags
, size
, p
);
3445 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3449 cache_alloc_debugcheck_after_bulk(s
, flags
, i
, p
, _RET_IP_
);
3450 slab_post_alloc_hook(s
, flags
, i
, p
);
3451 __kmem_cache_free_bulk(s
, i
, p
);
3454 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3456 #ifdef CONFIG_TRACING
3458 kmem_cache_alloc_trace(struct kmem_cache
*cachep
, gfp_t flags
, size_t size
)
3462 ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3464 kasan_kmalloc(cachep
, ret
, size
, flags
);
3465 trace_kmalloc(_RET_IP_
, ret
,
3466 size
, cachep
->size
, flags
);
3469 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3474 * kmem_cache_alloc_node - Allocate an object on the specified node
3475 * @cachep: The cache to allocate from.
3476 * @flags: See kmalloc().
3477 * @nodeid: node number of the target node.
3479 * Identical to kmem_cache_alloc but it will allocate memory on the given
3480 * node, which can improve the performance for cpu bound structures.
3482 * Fallback to other node is possible if __GFP_THISNODE is not set.
3484 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3486 void *ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3488 kasan_slab_alloc(cachep
, ret
, flags
);
3489 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3490 cachep
->object_size
, cachep
->size
,
3495 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3497 #ifdef CONFIG_TRACING
3498 void *kmem_cache_alloc_node_trace(struct kmem_cache
*cachep
,
3505 ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3507 kasan_kmalloc(cachep
, ret
, size
, flags
);
3508 trace_kmalloc_node(_RET_IP_
, ret
,
3513 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3516 static __always_inline
void *
3517 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, unsigned long caller
)
3519 struct kmem_cache
*cachep
;
3522 cachep
= kmalloc_slab(size
, flags
);
3523 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3525 ret
= kmem_cache_alloc_node_trace(cachep
, flags
, node
, size
);
3526 kasan_kmalloc(cachep
, ret
, size
, flags
);
3531 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3533 return __do_kmalloc_node(size
, flags
, node
, _RET_IP_
);
3535 EXPORT_SYMBOL(__kmalloc_node
);
3537 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3538 int node
, unsigned long caller
)
3540 return __do_kmalloc_node(size
, flags
, node
, caller
);
3542 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3543 #endif /* CONFIG_NUMA */
3546 * __do_kmalloc - allocate memory
3547 * @size: how many bytes of memory are required.
3548 * @flags: the type of memory to allocate (see kmalloc).
3549 * @caller: function caller for debug tracking of the caller
3551 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3552 unsigned long caller
)
3554 struct kmem_cache
*cachep
;
3557 cachep
= kmalloc_slab(size
, flags
);
3558 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3560 ret
= slab_alloc(cachep
, flags
, caller
);
3562 kasan_kmalloc(cachep
, ret
, size
, flags
);
3563 trace_kmalloc(caller
, ret
,
3564 size
, cachep
->size
, flags
);
3569 void *__kmalloc(size_t size
, gfp_t flags
)
3571 return __do_kmalloc(size
, flags
, _RET_IP_
);
3573 EXPORT_SYMBOL(__kmalloc
);
3575 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3577 return __do_kmalloc(size
, flags
, caller
);
3579 EXPORT_SYMBOL(__kmalloc_track_caller
);
3582 * kmem_cache_free - Deallocate an object
3583 * @cachep: The cache the allocation was from.
3584 * @objp: The previously allocated object.
3586 * Free an object which was previously allocated from this
3589 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3591 unsigned long flags
;
3592 cachep
= cache_from_obj(cachep
, objp
);
3596 local_irq_save(flags
);
3597 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3598 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3599 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3600 __cache_free(cachep
, objp
, _RET_IP_
);
3601 local_irq_restore(flags
);
3603 trace_kmem_cache_free(_RET_IP_
, objp
);
3605 EXPORT_SYMBOL(kmem_cache_free
);
3607 void kmem_cache_free_bulk(struct kmem_cache
*orig_s
, size_t size
, void **p
)
3609 struct kmem_cache
*s
;
3612 local_irq_disable();
3613 for (i
= 0; i
< size
; i
++) {
3616 if (!orig_s
) /* called via kfree_bulk */
3617 s
= virt_to_cache(objp
);
3619 s
= cache_from_obj(orig_s
, objp
);
3621 debug_check_no_locks_freed(objp
, s
->object_size
);
3622 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
3623 debug_check_no_obj_freed(objp
, s
->object_size
);
3625 __cache_free(s
, objp
, _RET_IP_
);
3629 /* FIXME: add tracing */
3631 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3634 * kfree - free previously allocated memory
3635 * @objp: pointer returned by kmalloc.
3637 * If @objp is NULL, no operation is performed.
3639 * Don't free memory not originally allocated by kmalloc()
3640 * or you will run into trouble.
3642 void kfree(const void *objp
)
3644 struct kmem_cache
*c
;
3645 unsigned long flags
;
3647 trace_kfree(_RET_IP_
, objp
);
3649 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3651 local_irq_save(flags
);
3652 kfree_debugcheck(objp
);
3653 c
= virt_to_cache(objp
);
3654 debug_check_no_locks_freed(objp
, c
->object_size
);
3656 debug_check_no_obj_freed(objp
, c
->object_size
);
3657 __cache_free(c
, (void *)objp
, _RET_IP_
);
3658 local_irq_restore(flags
);
3660 EXPORT_SYMBOL(kfree
);
3663 * This initializes kmem_cache_node or resizes various caches for all nodes.
3665 static int alloc_kmem_cache_node(struct kmem_cache
*cachep
, gfp_t gfp
)
3668 struct kmem_cache_node
*n
;
3669 struct array_cache
*new_shared
;
3670 struct alien_cache
**new_alien
= NULL
;
3672 for_each_online_node(node
) {
3674 if (use_alien_caches
) {
3675 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
3681 if (cachep
->shared
) {
3682 new_shared
= alloc_arraycache(node
,
3683 cachep
->shared
*cachep
->batchcount
,
3686 free_alien_cache(new_alien
);
3691 n
= get_node(cachep
, node
);
3693 struct array_cache
*shared
= n
->shared
;
3696 spin_lock_irq(&n
->list_lock
);
3699 free_block(cachep
, shared
->entry
,
3700 shared
->avail
, node
, &list
);
3702 n
->shared
= new_shared
;
3704 n
->alien
= new_alien
;
3707 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3708 cachep
->batchcount
+ cachep
->num
;
3709 spin_unlock_irq(&n
->list_lock
);
3710 slabs_destroy(cachep
, &list
);
3712 free_alien_cache(new_alien
);
3715 n
= kmalloc_node(sizeof(struct kmem_cache_node
), gfp
, node
);
3717 free_alien_cache(new_alien
);
3722 kmem_cache_node_init(n
);
3723 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
3724 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
3725 n
->shared
= new_shared
;
3726 n
->alien
= new_alien
;
3727 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3728 cachep
->batchcount
+ cachep
->num
;
3729 cachep
->node
[node
] = n
;
3734 if (!cachep
->list
.next
) {
3735 /* Cache is not active yet. Roll back what we did */
3738 n
= get_node(cachep
, node
);
3741 free_alien_cache(n
->alien
);
3743 cachep
->node
[node
] = NULL
;
3751 /* Always called with the slab_mutex held */
3752 static int __do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3753 int batchcount
, int shared
, gfp_t gfp
)
3755 struct array_cache __percpu
*cpu_cache
, *prev
;
3758 cpu_cache
= alloc_kmem_cache_cpus(cachep
, limit
, batchcount
);
3762 prev
= cachep
->cpu_cache
;
3763 cachep
->cpu_cache
= cpu_cache
;
3764 kick_all_cpus_sync();
3767 cachep
->batchcount
= batchcount
;
3768 cachep
->limit
= limit
;
3769 cachep
->shared
= shared
;
3774 for_each_online_cpu(cpu
) {
3777 struct kmem_cache_node
*n
;
3778 struct array_cache
*ac
= per_cpu_ptr(prev
, cpu
);
3780 node
= cpu_to_mem(cpu
);
3781 n
= get_node(cachep
, node
);
3782 spin_lock_irq(&n
->list_lock
);
3783 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
3784 spin_unlock_irq(&n
->list_lock
);
3785 slabs_destroy(cachep
, &list
);
3790 return alloc_kmem_cache_node(cachep
, gfp
);
3793 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3794 int batchcount
, int shared
, gfp_t gfp
)
3797 struct kmem_cache
*c
;
3799 ret
= __do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3801 if (slab_state
< FULL
)
3804 if ((ret
< 0) || !is_root_cache(cachep
))
3807 lockdep_assert_held(&slab_mutex
);
3808 for_each_memcg_cache(c
, cachep
) {
3809 /* return value determined by the root cache only */
3810 __do_tune_cpucache(c
, limit
, batchcount
, shared
, gfp
);
3816 /* Called with slab_mutex held always */
3817 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3824 if (!is_root_cache(cachep
)) {
3825 struct kmem_cache
*root
= memcg_root_cache(cachep
);
3826 limit
= root
->limit
;
3827 shared
= root
->shared
;
3828 batchcount
= root
->batchcount
;
3831 if (limit
&& shared
&& batchcount
)
3834 * The head array serves three purposes:
3835 * - create a LIFO ordering, i.e. return objects that are cache-warm
3836 * - reduce the number of spinlock operations.
3837 * - reduce the number of linked list operations on the slab and
3838 * bufctl chains: array operations are cheaper.
3839 * The numbers are guessed, we should auto-tune as described by
3842 if (cachep
->size
> 131072)
3844 else if (cachep
->size
> PAGE_SIZE
)
3846 else if (cachep
->size
> 1024)
3848 else if (cachep
->size
> 256)
3854 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3855 * allocation behaviour: Most allocs on one cpu, most free operations
3856 * on another cpu. For these cases, an efficient object passing between
3857 * cpus is necessary. This is provided by a shared array. The array
3858 * replaces Bonwick's magazine layer.
3859 * On uniprocessor, it's functionally equivalent (but less efficient)
3860 * to a larger limit. Thus disabled by default.
3863 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
3868 * With debugging enabled, large batchcount lead to excessively long
3869 * periods with disabled local interrupts. Limit the batchcount
3874 batchcount
= (limit
+ 1) / 2;
3876 err
= do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3878 pr_err("enable_cpucache failed for %s, error %d\n",
3879 cachep
->name
, -err
);
3884 * Drain an array if it contains any elements taking the node lock only if
3885 * necessary. Note that the node listlock also protects the array_cache
3886 * if drain_array() is used on the shared array.
3888 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
3889 struct array_cache
*ac
, int node
)
3893 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
3894 check_mutex_acquired();
3896 if (!ac
|| !ac
->avail
)
3904 spin_lock_irq(&n
->list_lock
);
3905 drain_array_locked(cachep
, ac
, node
, false, &list
);
3906 spin_unlock_irq(&n
->list_lock
);
3908 slabs_destroy(cachep
, &list
);
3912 * cache_reap - Reclaim memory from caches.
3913 * @w: work descriptor
3915 * Called from workqueue/eventd every few seconds.
3917 * - clear the per-cpu caches for this CPU.
3918 * - return freeable pages to the main free memory pool.
3920 * If we cannot acquire the cache chain mutex then just give up - we'll try
3921 * again on the next iteration.
3923 static void cache_reap(struct work_struct
*w
)
3925 struct kmem_cache
*searchp
;
3926 struct kmem_cache_node
*n
;
3927 int node
= numa_mem_id();
3928 struct delayed_work
*work
= to_delayed_work(w
);
3930 if (!mutex_trylock(&slab_mutex
))
3931 /* Give up. Setup the next iteration. */
3934 list_for_each_entry(searchp
, &slab_caches
, list
) {
3938 * We only take the node lock if absolutely necessary and we
3939 * have established with reasonable certainty that
3940 * we can do some work if the lock was obtained.
3942 n
= get_node(searchp
, node
);
3944 reap_alien(searchp
, n
);
3946 drain_array(searchp
, n
, cpu_cache_get(searchp
), node
);
3949 * These are racy checks but it does not matter
3950 * if we skip one check or scan twice.
3952 if (time_after(n
->next_reap
, jiffies
))
3955 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
;
3957 drain_array(searchp
, n
, n
->shared
, node
);
3959 if (n
->free_touched
)
3960 n
->free_touched
= 0;
3964 freed
= drain_freelist(searchp
, n
, (n
->free_limit
+
3965 5 * searchp
->num
- 1) / (5 * searchp
->num
));
3966 STATS_ADD_REAPED(searchp
, freed
);
3972 mutex_unlock(&slab_mutex
);
3975 /* Set up the next iteration */
3976 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_AC
));
3979 #ifdef CONFIG_SLABINFO
3980 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
3983 unsigned long active_objs
;
3984 unsigned long num_objs
;
3985 unsigned long active_slabs
= 0;
3986 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
3990 struct kmem_cache_node
*n
;
3994 for_each_kmem_cache_node(cachep
, node
, n
) {
3997 spin_lock_irq(&n
->list_lock
);
3999 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
4000 if (page
->active
!= cachep
->num
&& !error
)
4001 error
= "slabs_full accounting error";
4002 active_objs
+= cachep
->num
;
4005 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
4006 if (page
->active
== cachep
->num
&& !error
)
4007 error
= "slabs_partial accounting error";
4008 if (!page
->active
&& !error
)
4009 error
= "slabs_partial accounting error";
4010 active_objs
+= page
->active
;
4013 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
4014 if (page
->active
&& !error
)
4015 error
= "slabs_free accounting error";
4018 free_objects
+= n
->free_objects
;
4020 shared_avail
+= n
->shared
->avail
;
4022 spin_unlock_irq(&n
->list_lock
);
4024 num_slabs
+= active_slabs
;
4025 num_objs
= num_slabs
* cachep
->num
;
4026 if (num_objs
- active_objs
!= free_objects
&& !error
)
4027 error
= "free_objects accounting error";
4029 name
= cachep
->name
;
4031 pr_err("slab: cache %s error: %s\n", name
, error
);
4033 sinfo
->active_objs
= active_objs
;
4034 sinfo
->num_objs
= num_objs
;
4035 sinfo
->active_slabs
= active_slabs
;
4036 sinfo
->num_slabs
= num_slabs
;
4037 sinfo
->shared_avail
= shared_avail
;
4038 sinfo
->limit
= cachep
->limit
;
4039 sinfo
->batchcount
= cachep
->batchcount
;
4040 sinfo
->shared
= cachep
->shared
;
4041 sinfo
->objects_per_slab
= cachep
->num
;
4042 sinfo
->cache_order
= cachep
->gfporder
;
4045 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
4049 unsigned long high
= cachep
->high_mark
;
4050 unsigned long allocs
= cachep
->num_allocations
;
4051 unsigned long grown
= cachep
->grown
;
4052 unsigned long reaped
= cachep
->reaped
;
4053 unsigned long errors
= cachep
->errors
;
4054 unsigned long max_freeable
= cachep
->max_freeable
;
4055 unsigned long node_allocs
= cachep
->node_allocs
;
4056 unsigned long node_frees
= cachep
->node_frees
;
4057 unsigned long overflows
= cachep
->node_overflow
;
4059 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4060 allocs
, high
, grown
,
4061 reaped
, errors
, max_freeable
, node_allocs
,
4062 node_frees
, overflows
);
4066 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4067 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4068 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4069 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4071 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4072 allochit
, allocmiss
, freehit
, freemiss
);
4077 #define MAX_SLABINFO_WRITE 128
4079 * slabinfo_write - Tuning for the slab allocator
4081 * @buffer: user buffer
4082 * @count: data length
4085 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4086 size_t count
, loff_t
*ppos
)
4088 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4089 int limit
, batchcount
, shared
, res
;
4090 struct kmem_cache
*cachep
;
4092 if (count
> MAX_SLABINFO_WRITE
)
4094 if (copy_from_user(&kbuf
, buffer
, count
))
4096 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4098 tmp
= strchr(kbuf
, ' ');
4103 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4106 /* Find the cache in the chain of caches. */
4107 mutex_lock(&slab_mutex
);
4109 list_for_each_entry(cachep
, &slab_caches
, list
) {
4110 if (!strcmp(cachep
->name
, kbuf
)) {
4111 if (limit
< 1 || batchcount
< 1 ||
4112 batchcount
> limit
|| shared
< 0) {
4115 res
= do_tune_cpucache(cachep
, limit
,
4122 mutex_unlock(&slab_mutex
);
4128 #ifdef CONFIG_DEBUG_SLAB_LEAK
4130 static inline int add_caller(unsigned long *n
, unsigned long v
)
4140 unsigned long *q
= p
+ 2 * i
;
4154 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4160 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
,
4169 for (i
= 0, p
= page
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4172 for (j
= page
->active
; j
< c
->num
; j
++) {
4173 if (get_free_obj(page
, j
) == i
) {
4183 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4184 * mapping is established when actual object allocation and
4185 * we could mistakenly access the unmapped object in the cpu
4188 if (probe_kernel_read(&v
, dbg_userword(c
, p
), sizeof(v
)))
4191 if (!add_caller(n
, v
))
4196 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4198 #ifdef CONFIG_KALLSYMS
4199 unsigned long offset
, size
;
4200 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4202 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4203 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4205 seq_printf(m
, " [%s]", modname
);
4209 seq_printf(m
, "%p", (void *)address
);
4212 static int leaks_show(struct seq_file
*m
, void *p
)
4214 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4216 struct kmem_cache_node
*n
;
4218 unsigned long *x
= m
->private;
4222 if (!(cachep
->flags
& SLAB_STORE_USER
))
4224 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4228 * Set store_user_clean and start to grab stored user information
4229 * for all objects on this cache. If some alloc/free requests comes
4230 * during the processing, information would be wrong so restart
4234 set_store_user_clean(cachep
);
4235 drain_cpu_caches(cachep
);
4239 for_each_kmem_cache_node(cachep
, node
, n
) {
4242 spin_lock_irq(&n
->list_lock
);
4244 list_for_each_entry(page
, &n
->slabs_full
, lru
)
4245 handle_slab(x
, cachep
, page
);
4246 list_for_each_entry(page
, &n
->slabs_partial
, lru
)
4247 handle_slab(x
, cachep
, page
);
4248 spin_unlock_irq(&n
->list_lock
);
4250 } while (!is_store_user_clean(cachep
));
4252 name
= cachep
->name
;
4254 /* Increase the buffer size */
4255 mutex_unlock(&slab_mutex
);
4256 m
->private = kzalloc(x
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4258 /* Too bad, we are really out */
4260 mutex_lock(&slab_mutex
);
4263 *(unsigned long *)m
->private = x
[0] * 2;
4265 mutex_lock(&slab_mutex
);
4266 /* Now make sure this entry will be retried */
4270 for (i
= 0; i
< x
[1]; i
++) {
4271 seq_printf(m
, "%s: %lu ", name
, x
[2*i
+3]);
4272 show_symbol(m
, x
[2*i
+2]);
4279 static const struct seq_operations slabstats_op
= {
4280 .start
= slab_start
,
4286 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4290 n
= __seq_open_private(file
, &slabstats_op
, PAGE_SIZE
);
4294 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4299 static const struct file_operations proc_slabstats_operations
= {
4300 .open
= slabstats_open
,
4302 .llseek
= seq_lseek
,
4303 .release
= seq_release_private
,
4307 static int __init
slab_proc_init(void)
4309 #ifdef CONFIG_DEBUG_SLAB_LEAK
4310 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4314 module_init(slab_proc_init
);
4318 * ksize - get the actual amount of memory allocated for a given object
4319 * @objp: Pointer to the object
4321 * kmalloc may internally round up allocations and return more memory
4322 * than requested. ksize() can be used to determine the actual amount of
4323 * memory allocated. The caller may use this additional memory, even though
4324 * a smaller amount of memory was initially specified with the kmalloc call.
4325 * The caller must guarantee that objp points to a valid object previously
4326 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4327 * must not be freed during the duration of the call.
4329 size_t ksize(const void *objp
)
4334 if (unlikely(objp
== ZERO_SIZE_PTR
))
4337 size
= virt_to_cache(objp
)->object_size
;
4338 /* We assume that ksize callers could use the whole allocated area,
4339 * so we need to unpoison this area.
4341 kasan_krealloc(objp
, size
, GFP_NOWAIT
);
4345 EXPORT_SYMBOL(ksize
);